Object Detection Of Visible Components In Cervical Images

Cervical disease is a common high incidence disease among women.The diagnosis of cervical disease is to detect and analyze the visible components(cells,pathogenic microorganisms and miscellaneous bacteria)of the cervical image to find the lesions.Traditional method of detecting the visible components of cervical images completely depends on the manual reading of pathologists,and there are a large number of misdiagnosis and missed diagnosis cases.With the continuous development of in-depth learning,computer-aided diagnosis technology is widely used in the field of cervical diseases,providing accurate auxiliary diagnosis information for pathologists,and improving the accuracy of diagnosis of diseases such as precancerous lesions,cell atrophy,microbial infection and so on.Object detection of visible components in cervical images is a key technology for computer-aided diagnosis of cervical diseases.The existing object detection model can not meet the requirements of actual diagnosis.The challenges include: the size difference of pathogenic microorganisms is large,resulting in the detection of small pathogenic microorganisms;The labeling cost of cervical pathological images is high,and there is a lack of public data set of pathogenic microorganisms;The similarity between cervical cells increases the difficulty of cervical cell detection.In order to solve the above problems,the research contents of this paper are as follows:(1)Aiming at the problem of large difference in size and complex background of pathogenic microorganism,a multi-scale detection method of pathogenic microorganism based on recursive feature pyramid was proposed.In view of the similarity between the characteristics of some pathogenic microorganisms and normal cells,the residual mixed attention module was introduced in the feature extraction stage of YOLOv5 to enhance the significance of pathogenic microorganisms.In the feature fusion stage,the recursive feature pyramid and the detection layer of small objects are used to improve the detection rate of small objects.In the secondary feature extraction module,the deep separable convolution is used to replace the ordinary convolution,which reduces the number of parameters of the model and improves the performance.The experiment shows that the improved algorithm is superior to other networks in pathogenic microorganism data set,Pascal VOC2007,Pascal VOC2012 and CDetector public dataset.(2)Aiming at the problems of high labeling cost of pathogenic microorganisms and few public datasets,a semi-supervised detection method of pathogenic microorganisms based on active learning was proposed.First of all,the two models with different initial values of parameters are used as the teacher network for detection to improve the quality of false labels.Then the feature perturbation from the teacher model is introduced into the student model to enhance the generalization ability and robustness of the model.Then,according to the classification reliability and positioning coincidence degree of the image object prediction of the two teacher networks,the unlabeled samples with poor recognition ability were selected for experts to mark through the active learning strategy,and the labeled dataset was expanded.Finally,the circuit training model.Experiments show that this method can better mine the information of unlabeled samples and improve the accuracy of pathogenic microorganisms.(3)Aiming at the lack of intelligent diagnosis methods for cervical cell atrophy,an intelligent diagnosis method for cervical cell atrophy based on Circle Net cell detection was proposed.First of all,the Gaussian kernel radius generation method for generating heat maps by Circle Net is improved.The gated pyramid convolution module is designed.The pyramid convolution fully extracts the local and global information of the image,and the gated channel conversion effectively models the context information around the cell.The modified Circle Net model was used to detect the squamous epithelial cells in the surface,middle and basal layers.Then use SOLOV2 instance to split the nucleus.Finally,the total number ratio,the ratio of nucleus to cytoplasm,the degree of cell crowding and the ratio of cell number in each layer of the patch were calculated,and the index was input into the random forest classification model to grade the degree of atrophy.The experiment shows that this method has a high accuracy in the identification of the degree of cervical cell atrophy.In conclusion,this paper proposes effective solutions to the problems existing in object detection of cervical tangible components.These schemes combine the characteristics of different visible components in cervical images with the shortcomings of existing object detection models to optimize the model.Experiments show that the proposed method has good performance in the corresponding tasks.